3.3 The Multi-Objective Time Series Search Problem
3.3.3 A Randomized Bi-Objective Online Algorithm
In this section, a randomized algorithm for the bi-objective time series search problem is presented and analyzed by worst-component competitive analysis. As for the single-objective time series search problem, the competitive ratio is reduced by randomization.
Note that we consider the bi-objective time series search problem, i.e., n= 2. Addition-ally, we consider only the competitive ratio, rather than the strong competitive ratio.
Competitive randomized algorithms for the multi-objective time series search problem with n≥ 3 are subject to further research, as well as strongly competitive randomized algorithms for n≥2.
We follow the same approach as presented for the single-objective case in (El-Yaniv et al., 2001) and assume that Mi/mi = 2ki for ki ∈ N+, i = 1,2. Without loss of generality, suppose that k1 ≥k2. For the fluctuation ratio ϕof the multi-objective time series search problem we then haveϕ=(M1M2)/(m1m2)= 2k, wherek=k1+k2. Consider the deterministic algorithmsrpp-prodl for l= 0, . . . , k−1:
Algorithm 15:Multi-objective reservation price policy rpp-prodl.
1 fort= 1, . . . , T do
2 Acceptrt= p1t, p2t|
if p1t ·p2t ≥m1m22l.
The randomized algorithmexpo-multchooses strategyrpp-prodl,l= 0, . . . , k−1, with probability1/k each before the first request is revealed.
Theorem 3.3.8. For k1 ≥k2, the competitive ratio with respect tof1 of expo-multis given by
Rf1(expo-mult) = log(ϕ).
Proof. The idea for the worst case sequence presented to the online player by the adver-sary is analogous to the worst case sequence in the single-objective variant: The efficient offline solution for the adversary is supposed to be given by a request rj+1 = p1j, p2j| with p1t ·p2t just below the threshold m1m22j+1 for some j. This request will not be accepted by the online player when a policy rpp-prodl with l ≥ j+ 1 is selected. In order to prevent the online player from accepting this request in case a policyrpp-prodl
withl≤j is selected, the adversary initially presents requests ri = p1i, p2i|
that match the threshold for policyrpp-prodl, i.e., withp1i·p2i =m1m22i fori= 0, . . . , j. This way, the expected value of the request accepted by the online player is minimized. Choosing the index j which maximizes the competitive ratio then gives the worst case instance.
Note that, in order to maximize the competitive ratio (rather than the strong com-petitive ratio), the requests triggering the thresholds m1m22i with i ≤ j have to be chosen such that the request rj+1 remains the only efficient solution. Otherwise, the algorithm’s value would also be compared to the other efficient solutions leading to a smaller competitive ratio. See also Figures 3.5 and 3.6 for an illustration of the worst case instance.
Let rj+1 = p1j+1, p2j+1|
be an efficient offline solution and let j be an integer such that
m1m22j ≤p1j+1·p2j+1< m1m22j+1.
For the maximal choice of j = k, we have pij+1 = Mi, i = 1, . . . , n, due to bounded prices and m12k1 ·m22k2 =m1m22k = M1M2. This choice is not advantageous for the adversary, since the online player obtains this price with probability 1/k and any other choice forjgives the adversary the possibility to prevent the online player from obtaining the best price.
So, for any particular choice of j < k, the adversary selects p1j+1·p2j+1 arbitrarily close tom1m22j+1, since, thereby, the adversary’s return is increased whereas the return of the online player remains unchanged. Therefore,
p1j+1·p2j+1=m1m22j+1−, >0. (3.17) Ignoring the, efficient offline solutions are given by
rj+1= p1j+1, p2j+1|
=
xj+1, m1mx22j+1
j+1
|
for xj+1 feasible. (3.18) The feasible values forxj+1 are restricted byxj+1∈[m1, M1]and
m1m22j+1
xj+1 ∈[m2, M2] ⇔ m1m22j+1
M2 ≤xj+1 ≤m12j+1.
p1t p2t
m120m121m122 m123 m124 m125=M1
m220 m221 m222 m223=M2
p1t·p2t=m1m226 p1t·p2t=m1m227
rsr7
ǫ
rsr6
rsr5
rsr4
rsr3
rsr2
rsr1
Figure 3.5: Example for a worst case sequence in Case 1.1: k1 = 5, k2 = 3, j = 6, and x7=m124, the squares mark the requests.
Since M2=m22k2, we havexj+1 ∈ Ij+1 with Ij+1:=h
maxn
m1, m12j+1−k2o ,min
M1, m12j+1 i
For the following analysis, we consider three cases with respect to the relation of j,k1, and k2:
Case 1: j+ 1≥k1
By assumption, we havek1 ≥k2, and, due toj+1≥k1, we havej+1≥k2. Therefore, m12j+1−k2 ≥m1 and M1 =m12k1 ≤m12j+1. Thus, the intervalIj+1 is given by
Ij+1 =h
m12j+1−k2, M1i .
Due to (3.18), the competitive ratio with respect tof1(c) = maxi=1,...,nciis maximized by choosing the minimal or maximal feasible value forxj+1, i.e.,xj+1=m12j+1−k2 or xj+1 =M1:
Case 1.1: xj+1 =m12j+1−k2
Consider the minimal feasible value for xj+1, i.e., xj+1 = m12j+1−k2 (see r7 in Figure 3.5). Then, rj+1 is given by rj+1 = m12j+1−k2, m22k2|
. Request rj+1 will not be accepted by the online player when a policy rpp-prodl with l ≥ j+ 1 is selected. Note that for l = j+ 1, the request rj+1 is not accepted by the online player sincep1j+1·p2j+1=m1m22j+1−(see (3.17), we ignored theto improve the readability).
In order to prevent the online player from accepting this request rj+1 in case a policyrpp-prodlwith0≤l≤j is selected, the adversary initially presents requests ri = p1i, p2i|
with p1i ·p2i = m1m22i for i = 0, . . . , j (see also Figure 3.5). Thus, requestri is given by
ri = p1i, p2i|
=
xi, m1mx22i
i
|
, xi ∈ Ii.
The interval Ii of feasible values for xi is in this case determined by the following conditions: In order to obtain a feasible request, we need to guaranteexi ∈[m1, M1] and
m1m22i
xi ∈[m2, M2] ⇔ m12i−k2 ≤xi ≤m12i.
Furthermore, we have to ensure thatri 5rj+1, i.e., thatridoes not represent another efficient solution. Otherwise, the algorithm’s value would also be compared to this efficient solution leading to a smaller competitive ratio (see Definition 3.2.1). Hence, we havexi ≤m12j+1−k2 and
m1m22i
xi ≤m22k2 ⇔ xi ≥m12i−k2.
Since j+ 1 ≤ k = k1 +k2 and, hence, M1 = m12k1 ≥ m12j+1−k2, the interval of feasible valuesIi is given by
Ii =h maxn
m1, m12i−k2o
,minn
m12i, m12j+1−k2oi
. (3.19)
xj+1is chosen as theminimal feasible value in order to maximize the second compo-nentp2j+1 for the adversary. Thus, we are looking for themaximal feasible value for xi in order to minimize the second component p2i for the online player and, hence, maximize the competitive ratio with respect tof1.
Due to (3.19), we have xi ≤m12i for i= 0, . . . , j+ 1−k2 and xi ≤m12j+1−k2 for i=j+ 2−k2, . . . , j. Thus,ri is chosen as
ri = m12i, m2|
for i= 0, . . . , j+ 1−k2, and ri = m12j+1−k2, m22i−j−1+k2|
for i=j+ 2−k2, . . . , j.
For the expected value of thei-th component ofalgwith respect to its randomized
decisions, denoted by E[algi], we have
The competitive ratio with respect tof1 is then given by
k≥maxj+1≥k1
Note that the continuous maximum with respect tojis obtained forj=k−2+log(2)1 . However, for integer j≤k−1, the maximum is given by j=k−1.
Case 1.2: xj+1 =M1
Now, consider the maximal value for xj+1, i.e., xj+1 = M1 (see r7 in Figure 3.6).
Then,rj+1 is given byrj+1 = M1, m22j+1−k1|
. By the same argumentation as in the previous case, the adversary previously presents requests
ri = p1i, p2i|
where Ii is again the set of feasible values for xi (see also Figure 3.6). In order to obtain a feasible request, we need to guaranteexi∈[m1, M1]and
p1t feasible valuesIi is given by
Ii =h
Sincexj+1 is in this case chosen as themaximal feasible value in order to maximize the first component p1j+1 for the adversary, we are looking for theminimal feasible value for xi in order to minimize the first component p1i for the online player and, hence, maximize the competitive ratio with respect tof1. Thus,ri is given by
ri = m1, m22i|
for i= 0, . . . , j+ 1−k1, and ri = m12i−j−1+k1, m22j+1−k1|
for i=j+ 2−k1, . . . , j.
For the expected value of alg we then have E[alg1] = 1
= m2 k
2j−k1+2−1 + (k1−1) 2j+1−k1 +k−j−1
= m2 k
(k1+ 1)2j+1−k1+k−j−2 . The competitive ratio with respect tof1 is then given by
k≥maxj+1≥k1
max
M1
E[alg1], m22j+1−k1 E[alg2]
= max
k≥j+1≥k1
max
k2k1
2k1+k2−1, k2j+1−k1
(k1+ 1)2j+1−k1 +k−j−2
(j+1=k)
= max
( log(ϕ) 1 +k2−1
2k1
, log(ϕ) k1+ 1− 21k2
)
. (3.22)
Note that the competitive ratio is symmetrical to the competitive ratio in Case 1.1 with respect to k1 andk2.
Case 2: j+ 1≤k2
By assumption, we havek1 ≥k2, and, due toj+1≤k2, we havej+1≤k1. Therefore, m1 ≥m12j+1−k2 and m12j+1 ≤m12k1 =M1. Thus, the intervalIj+1 is given by
Ij+1 =
m1, m12j+1 .
Due to (3.18), the competitive ratio with respect to f1 is maximized by choosing the minimal or maximal feasible value for xj+1, i.e., xj+1 = m1 or xj+1 = m12j+1. We proceed with the analysis as in Case 1:
Case 2.1: xj+1 =m1
Consider the minimal feasible value for xj+1, i.e., xj+1 = m1. Then, rj+1 is given by rj+1= m1, m22j+1|
. By the same argumentation as in the previous cases, the adversary previously presents requests
ri = p1i, p2i|
=
xi, m1xm22i
i
|
, xi∈ Ii.
The intervalIiis determined byxi∈[m1, M1],m12i−k2 ≤xi ≤m12i, andri5rj+1, i.e., m12i−j−1 ≤xi ≤m1 (for a detailed explanation see the previous cases). Conse-quently, Ii is given byIi = [m1, m1], i.e., xi =m1 and we have
ri = m1, m22i|
for i= 0, . . . , j.
For the expected value of alg we then have The competitive ratio with respect tof1 is then given by
j+1max≤k2
. As seen in Case 1.1 and Case 1.2, this case is symmetrical with respect tok1 andk2 to Case 2.1. Consequently, we have for the expected value of alg
E[alg1] = m1
k 2j+1+k−j−2 , E[alg2] =m2,
and the competitive ratio with respect tof1 is given by
j+1≤kmax2
Due to (3.18), the competitive ratio with respect tof1 is maximized by choosing the minimal or maximal feasible value forxj+1, i.e., xj+1 =m12j+1−k2 or xj+1 =m12j+1. These cases have been analyzed before, see Case 1.1 and Case 2.2.
Now, we bring together the competitive ratios of the different cases analyzed above.
By (3.20), (3.22), (3.23), and (3.24), the competitive ratio with respect to f1 of expo-mult is given by
Rf1(expo-mult) = max
( log(ϕ) k2+ 1−21k1
, log(ϕ) 1 +k1−1
2k2
, log(ϕ) 1 +k2−1
2k1
, log(ϕ) k1+ 1− 21k2
) . Then, we have
1 +k2−1
2k1 ≤k2+ 1− 1
2k1 and 1 +k1−1
2k2 ≤k1+ 1− 1 2k2, and, due to k1 ≥k2,
k2−1
2k1 ≤ k1−1 2k2 . Therefore, we have
Rf1(expo-mult) = log(ϕ) 1 +k2−1
2k1
≤log(ϕ).